Thinking Machines ships Inkling 2 Mini: looped-layer MoE for local use, rivals Qwen 3.8 27B

pcuenq · x · 2026-11-21

Hugging Face welcomes Thinking Machines Lab's Inkling 2 Mini, a looped transformer MoE with 26B total and 3B active parameters per token—the middle 12 of 24 decoder layers loop three times, acting like 48 logical layers. It supports 1M-token context, image+text input, ships with NVFP4/MXFP8 expert weights and a DSpark drafter for speedup. HF's vibe tests find it at least comparable to Qwen 3.8 27B for local setups. Day-zero support in transformers, llama.cpp, MLX, and SGLang.

Related event: Thinking Machines Releases 26B Looped Transformer Inkling 2 Mini, Runs on Laptops(115 posts)→

Original post →

More from Models

Models channel →